xref: /linux/net/ipv4/inet_timewait_sock.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * INET		An implementation of the TCP/IP protocol suite for the LINUX
4  *		operating system.  INET is implemented using the  BSD Socket
5  *		interface as the means of communication with the user level.
6  *
7  *		Generic TIME_WAIT sockets functions
8  *
9  *		From code orinally in TCP
10  */
11 
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/module.h>
15 #include <net/inet_hashtables.h>
16 #include <net/inet_timewait_sock.h>
17 #include <net/ip.h>
18 
19 
20 /**
21  *	inet_twsk_bind_unhash - unhash a timewait socket from bind hash
22  *	@tw: timewait socket
23  *	@hashinfo: hashinfo pointer
24  *
25  *	unhash a timewait socket from bind hash, if hashed.
26  *	bind hash lock must be held by caller.
27  *	Returns 1 if caller should call inet_twsk_put() after lock release.
28  */
29 void inet_twsk_bind_unhash(struct inet_timewait_sock *tw,
30 			  struct inet_hashinfo *hashinfo)
31 {
32 	struct inet_bind2_bucket *tb2 = tw->tw_tb2;
33 	struct inet_bind_bucket *tb = tw->tw_tb;
34 
35 	if (!tb)
36 		return;
37 
38 	__sk_del_bind_node((struct sock *)tw);
39 	tw->tw_tb = NULL;
40 	tw->tw_tb2 = NULL;
41 	inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
42 	inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
43 
44 	__sock_put((struct sock *)tw);
45 }
46 
47 /* Must be called with locally disabled BHs. */
48 static void inet_twsk_kill(struct inet_timewait_sock *tw)
49 {
50 	struct inet_hashinfo *hashinfo = tw->tw_dr->hashinfo;
51 	spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
52 	struct inet_bind_hashbucket *bhead, *bhead2;
53 
54 	spin_lock(lock);
55 	sk_nulls_del_node_init_rcu((struct sock *)tw);
56 	spin_unlock(lock);
57 
58 	/* Disassociate with bind bucket. */
59 	bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), tw->tw_num,
60 			hashinfo->bhash_size)];
61 	bhead2 = inet_bhashfn_portaddr(hashinfo, (struct sock *)tw,
62 				       twsk_net(tw), tw->tw_num);
63 
64 	spin_lock(&bhead->lock);
65 	spin_lock(&bhead2->lock);
66 	inet_twsk_bind_unhash(tw, hashinfo);
67 	spin_unlock(&bhead2->lock);
68 	spin_unlock(&bhead->lock);
69 
70 	refcount_dec(&tw->tw_dr->tw_refcount);
71 	inet_twsk_put(tw);
72 }
73 
74 void inet_twsk_free(struct inet_timewait_sock *tw)
75 {
76 	struct module *owner = tw->tw_prot->owner;
77 	twsk_destructor((struct sock *)tw);
78 	kmem_cache_free(tw->tw_prot->twsk_prot->twsk_slab, tw);
79 	module_put(owner);
80 }
81 
82 void inet_twsk_put(struct inet_timewait_sock *tw)
83 {
84 	if (refcount_dec_and_test(&tw->tw_refcnt))
85 		inet_twsk_free(tw);
86 }
87 EXPORT_SYMBOL_GPL(inet_twsk_put);
88 
89 static void inet_twsk_add_node_rcu(struct inet_timewait_sock *tw,
90 				   struct hlist_nulls_head *list)
91 {
92 	hlist_nulls_add_head_rcu(&tw->tw_node, list);
93 }
94 
95 static void inet_twsk_schedule(struct inet_timewait_sock *tw, int timeo)
96 {
97 	__inet_twsk_schedule(tw, timeo, false);
98 }
99 
100 /*
101  * Enter the time wait state.
102  * Essentially we whip up a timewait bucket, copy the relevant info into it
103  * from the SK, and mess with hash chains and list linkage.
104  *
105  * The caller must not access @tw anymore after this function returns.
106  */
107 void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,
108 				  struct sock *sk,
109 				  struct inet_hashinfo *hashinfo,
110 				  int timeo)
111 {
112 	const struct inet_sock *inet = inet_sk(sk);
113 	const struct inet_connection_sock *icsk = inet_csk(sk);
114 	struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
115 	spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
116 	struct inet_bind_hashbucket *bhead, *bhead2;
117 
118 	/* Step 1: Put TW into bind hash. Original socket stays there too.
119 	   Note, that any socket with inet->num != 0 MUST be bound in
120 	   binding cache, even if it is closed.
121 	 */
122 	bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), inet->inet_num,
123 			hashinfo->bhash_size)];
124 	bhead2 = inet_bhashfn_portaddr(hashinfo, sk, twsk_net(tw), inet->inet_num);
125 
126 	local_bh_disable();
127 	spin_lock(&bhead->lock);
128 	spin_lock(&bhead2->lock);
129 
130 	tw->tw_tb = icsk->icsk_bind_hash;
131 	WARN_ON(!icsk->icsk_bind_hash);
132 
133 	tw->tw_tb2 = icsk->icsk_bind2_hash;
134 	WARN_ON(!icsk->icsk_bind2_hash);
135 	sk_add_bind_node((struct sock *)tw, &tw->tw_tb2->owners);
136 
137 	spin_unlock(&bhead2->lock);
138 	spin_unlock(&bhead->lock);
139 
140 	spin_lock(lock);
141 
142 	/* Step 2: Hash TW into tcp ehash chain */
143 	inet_twsk_add_node_rcu(tw, &ehead->chain);
144 
145 	/* Step 3: Remove SK from hash chain */
146 	if (__sk_nulls_del_node_init_rcu(sk))
147 		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
148 
149 
150 	/* Ensure above writes are committed into memory before updating the
151 	 * refcount.
152 	 * Provides ordering vs later refcount_inc().
153 	 */
154 	smp_wmb();
155 	/* tw_refcnt is set to 3 because we have :
156 	 * - one reference for bhash chain.
157 	 * - one reference for ehash chain.
158 	 * - one reference for timer.
159 	 * Also note that after this point, we lost our implicit reference
160 	 * so we are not allowed to use tw anymore.
161 	 */
162 	refcount_set(&tw->tw_refcnt, 3);
163 
164 	inet_twsk_schedule(tw, timeo);
165 
166 	spin_unlock(lock);
167 	local_bh_enable();
168 }
169 EXPORT_SYMBOL_GPL(inet_twsk_hashdance_schedule);
170 
171 static void tw_timer_handler(struct timer_list *t)
172 {
173 	struct inet_timewait_sock *tw = from_timer(tw, t, tw_timer);
174 
175 	inet_twsk_kill(tw);
176 }
177 
178 struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk,
179 					   struct inet_timewait_death_row *dr,
180 					   const int state)
181 {
182 	struct inet_timewait_sock *tw;
183 
184 	if (refcount_read(&dr->tw_refcount) - 1 >=
185 	    READ_ONCE(dr->sysctl_max_tw_buckets))
186 		return NULL;
187 
188 	tw = kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
189 			      GFP_ATOMIC);
190 	if (tw) {
191 		const struct inet_sock *inet = inet_sk(sk);
192 
193 		tw->tw_dr	    = dr;
194 		/* Give us an identity. */
195 		tw->tw_daddr	    = inet->inet_daddr;
196 		tw->tw_rcv_saddr    = inet->inet_rcv_saddr;
197 		tw->tw_bound_dev_if = sk->sk_bound_dev_if;
198 		tw->tw_tos	    = inet->tos;
199 		tw->tw_num	    = inet->inet_num;
200 		tw->tw_state	    = TCP_TIME_WAIT;
201 		tw->tw_substate	    = state;
202 		tw->tw_sport	    = inet->inet_sport;
203 		tw->tw_dport	    = inet->inet_dport;
204 		tw->tw_family	    = sk->sk_family;
205 		tw->tw_reuse	    = sk->sk_reuse;
206 		tw->tw_reuseport    = sk->sk_reuseport;
207 		tw->tw_hash	    = sk->sk_hash;
208 		tw->tw_ipv6only	    = 0;
209 		tw->tw_transparent  = inet_test_bit(TRANSPARENT, sk);
210 		tw->tw_prot	    = sk->sk_prot_creator;
211 		atomic64_set(&tw->tw_cookie, atomic64_read(&sk->sk_cookie));
212 		twsk_net_set(tw, sock_net(sk));
213 		timer_setup(&tw->tw_timer, tw_timer_handler, 0);
214 		/*
215 		 * Because we use RCU lookups, we should not set tw_refcnt
216 		 * to a non null value before everything is setup for this
217 		 * timewait socket.
218 		 */
219 		refcount_set(&tw->tw_refcnt, 0);
220 
221 		__module_get(tw->tw_prot->owner);
222 	}
223 
224 	return tw;
225 }
226 EXPORT_SYMBOL_GPL(inet_twsk_alloc);
227 
228 /* These are always called from BH context.  See callers in
229  * tcp_input.c to verify this.
230  */
231 
232 /* This is for handling early-kills of TIME_WAIT sockets.
233  * Warning : consume reference.
234  * Caller should not access tw anymore.
235  */
236 void inet_twsk_deschedule_put(struct inet_timewait_sock *tw)
237 {
238 	struct inet_hashinfo *hashinfo = tw->tw_dr->hashinfo;
239 	spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
240 
241 	/* inet_twsk_purge() walks over all sockets, including tw ones,
242 	 * and removes them via inet_twsk_deschedule_put() after a
243 	 * refcount_inc_not_zero().
244 	 *
245 	 * inet_twsk_hashdance_schedule() must (re)init the refcount before
246 	 * arming the timer, i.e. inet_twsk_purge can obtain a reference to
247 	 * a twsk that did not yet schedule the timer.
248 	 *
249 	 * The ehash lock synchronizes these two:
250 	 * After acquiring the lock, the timer is always scheduled (else
251 	 * timer_shutdown returns false), because hashdance_schedule releases
252 	 * the ehash lock only after completing the timer initialization.
253 	 *
254 	 * Without grabbing the ehash lock, we get:
255 	 * 1) cpu x sets twsk refcount to 3
256 	 * 2) cpu y bumps refcount to 4
257 	 * 3) cpu y calls inet_twsk_deschedule_put() and shuts timer down
258 	 * 4) cpu x tries to start timer, but mod_timer is a noop post-shutdown
259 	 * -> timer refcount is never decremented.
260 	 */
261 	spin_lock(lock);
262 	/*  Makes sure hashdance_schedule() has completed */
263 	spin_unlock(lock);
264 
265 	if (timer_shutdown_sync(&tw->tw_timer))
266 		inet_twsk_kill(tw);
267 	inet_twsk_put(tw);
268 }
269 EXPORT_SYMBOL(inet_twsk_deschedule_put);
270 
271 void __inet_twsk_schedule(struct inet_timewait_sock *tw, int timeo, bool rearm)
272 {
273 	/* timeout := RTO * 3.5
274 	 *
275 	 * 3.5 = 1+2+0.5 to wait for two retransmits.
276 	 *
277 	 * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
278 	 * our ACK acking that FIN can be lost. If N subsequent retransmitted
279 	 * FINs (or previous seqments) are lost (probability of such event
280 	 * is p^(N+1), where p is probability to lose single packet and
281 	 * time to detect the loss is about RTO*(2^N - 1) with exponential
282 	 * backoff). Normal timewait length is calculated so, that we
283 	 * waited at least for one retransmitted FIN (maximal RTO is 120sec).
284 	 * [ BTW Linux. following BSD, violates this requirement waiting
285 	 *   only for 60sec, we should wait at least for 240 secs.
286 	 *   Well, 240 consumes too much of resources 8)
287 	 * ]
288 	 * This interval is not reduced to catch old duplicate and
289 	 * responces to our wandering segments living for two MSLs.
290 	 * However, if we use PAWS to detect
291 	 * old duplicates, we can reduce the interval to bounds required
292 	 * by RTO, rather than MSL. So, if peer understands PAWS, we
293 	 * kill tw bucket after 3.5*RTO (it is important that this number
294 	 * is greater than TS tick!) and detect old duplicates with help
295 	 * of PAWS.
296 	 */
297 
298 	if (!rearm) {
299 		bool kill = timeo <= 4*HZ;
300 
301 		__NET_INC_STATS(twsk_net(tw), kill ? LINUX_MIB_TIMEWAITKILLED :
302 						     LINUX_MIB_TIMEWAITED);
303 		BUG_ON(mod_timer(&tw->tw_timer, jiffies + timeo));
304 		refcount_inc(&tw->tw_dr->tw_refcount);
305 	} else {
306 		mod_timer_pending(&tw->tw_timer, jiffies + timeo);
307 	}
308 }
309 EXPORT_SYMBOL_GPL(__inet_twsk_schedule);
310 
311 /* Remove all non full sockets (TIME_WAIT and NEW_SYN_RECV) for dead netns */
312 void inet_twsk_purge(struct inet_hashinfo *hashinfo)
313 {
314 	struct inet_ehash_bucket *head = &hashinfo->ehash[0];
315 	unsigned int ehash_mask = hashinfo->ehash_mask;
316 	struct hlist_nulls_node *node;
317 	unsigned int slot;
318 	struct sock *sk;
319 
320 	for (slot = 0; slot <= ehash_mask; slot++, head++) {
321 		if (hlist_nulls_empty(&head->chain))
322 			continue;
323 
324 restart_rcu:
325 		cond_resched();
326 		rcu_read_lock();
327 restart:
328 		sk_nulls_for_each_rcu(sk, node, &head->chain) {
329 			int state = inet_sk_state_load(sk);
330 
331 			if ((1 << state) & ~(TCPF_TIME_WAIT |
332 					     TCPF_NEW_SYN_RECV))
333 				continue;
334 
335 			if (refcount_read(&sock_net(sk)->ns.count))
336 				continue;
337 
338 			if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
339 				continue;
340 
341 			if (refcount_read(&sock_net(sk)->ns.count)) {
342 				sock_gen_put(sk);
343 				goto restart;
344 			}
345 
346 			rcu_read_unlock();
347 			local_bh_disable();
348 			if (state == TCP_TIME_WAIT) {
349 				inet_twsk_deschedule_put(inet_twsk(sk));
350 			} else {
351 				struct request_sock *req = inet_reqsk(sk);
352 
353 				inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
354 								  req);
355 			}
356 			local_bh_enable();
357 			goto restart_rcu;
358 		}
359 		/* If the nulls value we got at the end of this lookup is
360 		 * not the expected one, we must restart lookup.
361 		 * We probably met an item that was moved to another chain.
362 		 */
363 		if (get_nulls_value(node) != slot)
364 			goto restart;
365 		rcu_read_unlock();
366 	}
367 }
368 EXPORT_SYMBOL_GPL(inet_twsk_purge);
369